XLA debug flags

xla
1

Hi,

Is it possible to pass XLA debug flags to torch-xla?

E.g., when using jax, setting XLA_FLAGS=--xla_force_host_platform_device_count=8 will mimic an 8 device / core backend (useful for debugging in CPU-only platforms) such that jax.devices() will return a list of 8 devices.

An equivalent call to torch_xla.xla_model.get_xla_supported_devices() would have similar functionality.

2

Hi Jerome,

It is possible to do that through env var. One example is the flag we passed to xla directly xla/__init__.py at master · pytorch/xla · GitHub.

BTW we checked the github issue more often so next time you have issues feel free to open issues on Issues · pytorch/xla · GitHub.